Hybrid noise suppression for communication systems

Abstract

A method for hybrid noise suppression includes receiving a processed audio signal from an audio device. The processed audio signal results from a partial audio processing performed on a noisy audio input signal. The method further includes predicting a noise suppression parameter using a neural network model operating on the processed audio signal and generating a noise-suppressed audio signal from the processed audio signal, using the noise suppression parameter. The method further includes generating a noise-suppressed audio output signal from the noise-suppressed audio signal using an additional audio processing and outputting the noise-suppressed audio output signal.

What is claimed is: 
     
       1. A method for hybrid noise suppression, comprising:
 receiving a processed audio signal from an audio device,
 wherein the processed audio signal results from a partial audio processing performed on a noisy audio input signal; 
 
 predicting a noise suppression parameter using a neural network model operating on the processed audio signal; 
 generating a noise-suppressed audio signal from the processed audio signal, using the noise suppression parameter; 
 generating a noise-suppressed audio output signal from the noise-suppressed audio signal using an additional audio processing, wherein the additional audio processing comprises a noise spectrum estimate-based noise suppression; and 
 outputting the noise-suppressed audio output signal. 
 
     
     
       2. The method of  claim 1 , wherein predicting the noise suppression parameter comprises:
 transforming the processed audio signal from a time domain into a frequency domain in a plurality of frequency sub-bands, 
 generating a feature vector from the processed audio signal in the plurality of frequency sub-bands, and 
 obtaining a set of sub-band gain values by applying the neural network model to the feature vector, wherein the set of sub-band gain values forms the noise suppression parameter. 
 
     
     
       3. The method of  claim 1 , wherein generating the noise-suppressed audio signal comprises:
 transforming the processed audio signal from a time domain into a frequency domain in a plurality of frequency sub-bands, and 
 scaling, according to sub-band gain values provided as the noise suppression parameter, the processed audio signal in the plurality of frequency sub-bands to generate the noise-suppressed audio signal. 
 
     
     
       4. The method of  claim 3 , wherein generating the noise-suppressed audio signal further comprises:
 transforming the noise-suppressed audio signal from the frequency domain to the time domain. 
 
     
     
       5. The method of  claim 1 , wherein the neural network model is a deep neural network. 
     
     
       6. The method of  claim 5 , wherein the deep neural network comprises gated recurrent units. 
     
     
       7. The method of  claim 1 , wherein the noise spectrum estimate-based noise suppression comprises:
 transforming the noise-suppressed audio signal from a time domain into a frequency domain comprising a plurality of frequency sub-bands, 
 obtaining a noise estimate in the plurality of frequency sub-bands, 
 obtaining the noise-suppressed audio output signal by removing the noise estimate from the noise-suppressed audio signal for the plurality of frequency sub-bands, and 
 transforming, after obtaining the noise-suppressed audio output signal, the noise-suppressed audio output signal from the frequency domain to the time domain. 
 
     
     
       8. The method of  claim 1 , wherein the additional audio processing further comprises an automatic gain control. 
     
     
       9. The method of  claim 1 , wherein the additional audio processing further comprises an equalizing. 
     
     
       10. The method of  claim 1 , wherein the noise-suppressed audio output signal is in a set of frequency sub-bands, and wherein the noise-suppressed audio output signal is transformed from a frequency domain to a time domain prior to outputting. 
     
     
       11. The method of  claim 1 , wherein the partial audio processing comprises at least one selected from the group consisting of a beamforming, an automatic gain control, an equalizing, an echo cancellation, and a limiting. 
     
     
       12. A system for hybrid noise suppression, comprising a host comprising:
 a memory; and 
 circuitry for performing operations using the memory, the operations comprising:
 receiving a processed audio signal from an audio device,
 wherein the processed audio signal results from a partial audio processing performed on a noisy audio input signal; 
 
 predicting a noise suppression parameter using a neural network model operating on the processed audio signal; 
 generating a noise-suppressed audio signal from the processed audio signal, using the noise suppression parameter; 
 generating a noise-suppressed audio output signal from the noise-suppressed audio signal using an additional audio processing, wherein the additional audio processing comprises a noise spectrum estimate-based noise suppression; and 
 outputting the noise-suppressed audio output signal. 
 
 
     
     
       13. The system of  claim 12 , wherein predicting the noise suppression parameter comprises:
 transforming the processed audio signal from a time domain into a frequency domain in a plurality of frequency sub-bands, 
 generating a feature vector from the processed audio signal in the plurality of frequency sub-bands, and 
 obtaining a set of sub-band gain values by applying the neural network model to the feature vector, wherein the set of sub-band gain values forms the noise suppression parameter. 
 
     
     
       14. The system of  claim 12 , wherein generating the noise-suppressed audio signal comprises:
 transforming the processed audio signal from a time domain into a frequency domain in a plurality of frequency sub-bands, and 
 scaling, according to sub-band gain values provided as the noise suppression parameter, the processed audio signal in the plurality of frequency sub-bands to generate the noise-suppressed audio signal. 
 
     
     
       15. The system of  claim 14 , wherein generating the noise-suppressed audio signal further comprises:
 transforming the noise-suppressed audio signal from the frequency domain to the time domain. 
 
     
     
       16. The system of  claim 12 , wherein the neural network model is a deep neural network. 
     
     
       17. The system of  claim 12 , wherein the noise spectrum estimate-based noise suppression comprises:
 transforming the noise-suppressed audio signal from a time domain into a frequency domain comprising a plurality of frequency sub-bands, 
 obtaining a noise estimate in the plurality of frequency sub-bands, 
 obtaining the noise-suppressed audio output signal by removing the noise estimate from the noise-suppressed audio signal for the plurality of frequency sub-bands, and 
 transforming, after obtaining the noise-suppressed audio output signal, the noise-suppressed audio output signal from the frequency domain to the time domain. 
 
     
     
       18. A system for hybrid noise suppression, comprising:
 an audio device for performing operations comprising:
 performing partial audio processing on a noisy audio input signal, and 
 outputting the processed audio signal results; and 
 
 a host for performing operations comprising:
 receiving the processed audio signal from the audio device, 
 predicting a noise suppression parameter using a neural network model operating on the processed audio signal, 
 generating a noise-suppressed audio signal from the processed audio signal, using the noise suppression parameter, 
 generating a noise-suppressed audio output signal from the noise-suppressed audio signal using an additional audio processing, wherein the additional audio processing comprises a noise spectrum estimate-based noise suppression, and 
 outputting the noise-suppressed audio output signal.